coffeemug

I disagree with the first part (that this is merely a voice). There is a distinct difference between an author's unique voice and slop. It may be hard to tease out exactly what the difference is, but it seems self-evident to me it's there. (I'll need to think more how to make the distinction explicit; it's not immediately obvious to me how to discriminate between the two.)

EDIT: ok, here are two ways:

1. if it's merely a voice, I want to hear it. If it's slop, I want it taken out.

2. voice is signal, slop is noise; thus low-signal sentences are slop.

Would you feel the same way if he attached good faith probability estimates to those claims?

As a language Lisp is great (though the ecosystem is limited). It has two flaws-- it's very open-ended so unless you're talented and disciplined you can fall into a rabbit hole of hacking fun Lisp stuff and not actually getting any work done. Other languages have this problem, but Lisp I think more so.

The second flaw is that it eats perfectly capable minds for years and the results don't justify the time investment. Python or whatever is fine. I wish I took most (but not all) of the time I invested in Lisp and put it into something else instead.

I don’t find Opus to be overly left-leaning. I think it’s generally pretty balanced with a slight lean left, but when pressed it will happily steelman rightwing arguments and operate within a right-leaning belief framework in good faith. (I found OpenAI’s models much less willing to do that, but it’s been a while since I tried, I should retest.)

Not anything. There are domains where precedents empirically don’t balloon and domains where they do. Federal government is nearly always the latter.

The issues with this are precedent/slippery slope. Today it's a small stake in a small number of companies, but empirically government initiatives almost always grow. Once entrenched, I would be willing to bet that the government will take larger and larger stakes in a bigger and bigger number of companies until this model becomes a non-trivial, potentially dominant, share of the economy. I can see myself becoming a single-issue voter to oppose that future (although given bipartisan support I'm unlikely to have any options; maybe Rand Paul).

When I was lifting weights I went from 170lbs to 210, and back to 170. Doing that isn’t technically hard, and maintaining is even easier. But I had to think about it _all_ the time. Maybe it’s because I’m older now, but I just don’t want to spend that much effort thinking about food.

Regulatory agencies limit uses of other products without acts of congress-- cigarettes, vapes, drugs, pesticides, chemicals, explosives. Even firearms, despite a constitutional amendment! Why not models? (Note I am not arguing it's a good idea; I'm making a narrow argument that there is precedent.)

EDIT: I agree that it should require an act of Congress to explicitly delegate this power.

Another plausible future is AI reshuffling the economic hierarchy. In a technological civilization the pressure valve need not be political.

As an aside, you made me curious if Trump made this constituency materially better off. Here's what Claude thinks (tl;dr: it's a wash): https://claude.ai/share/36233694-3729-4758-b2e6-c2058791ab1a

It depends on whether you believe US action is overdetermined, but I think if Trump didn’t get elected we would have continued on the path of free trade. His election wasn’t predestined. He had just the right mix of features to win at the time, but if this basket of features didn’t exist it’s not hard to imagine the country going down a very different path.

That’s a very uncharitable way to interpret public opinion. A more charitable way is that people have different ethical intuitions. Some of us lean more utilitarian than others, and almost no one is a pure utilitarian. It isn’t merely a matter of being more emotional; people very reasonably have different views on this stuff.

“Billionaires v. the underclass” is such an unsophisticated discussion. Markets have all kinds of externalities, equilibrium traps, information asymmetry flywheels, and hundreds of other phenomena that interplay to trip people’s sense of fairness in all kinds of ways.

To pick just one example, infinite scrolling can be seen as a modern equivalent of cigarettes— a product that made people billionaires, and that consumers obviously want but are not free to stop using because of hyper-sophisticated dark patterns.

Is Elon a trillionaire because he created a trillion dollars of value from thin air, or is it because he created an information asymmetry flywheel that lets him allocate capital more efficiently than other actors?

It’s genuinely unclear to me whether the universe in which we incentivize this kind of scale is better than the universe in which we do not (because the counterfactual universe has massive externalities too). But this is obviously not just a matter of compounding value creation and becoming a billionaire fair and square in ten years.

Giving normal people something that has only been available to rich people is a staple of technological innovation. The problem in this case with Siri isn’t that people don’t want an assistant. It’s that it doesn’t actually work yet.

Microsoft is no longer rest and vest and hasn’t been for a long time. The expectations and pressure are very high (at least the corner I’m in)

The balance of power between capital and labor fluctuates; qualitatively it definitely felt different ten years ago. For whatever reason labor seems to have much less power today. Not zero, just less. It isn't that companies used to be more loyal out of some moral obligation. They were forced to be more loyal by market dynamics that used to be more favorable to labor.

Both arguments can be (and are) right at once.

What OP said is definitely true on the micro level-- not "even/might/some aspects", but the whole thing. It's true that in any given organization there are fewer senior roles because of hierarchical nature, it's true that as you progress up the ladder the demands change and increase, and it's true that many people fail or choose not to adapt.

The macro argument seems right as well. If you measure it longitudinally the numbers don't stay constant. It's 1 in 4 today, maybe it was 1 in 10 fifteen years ago. Anecdotally there is definitely something strange going on with the labor market that's new, and that you can't explain by micro realities alone.

I think your first equilibrium would be hard, for two reasons. First, empirically insurgencies are extremely difficult to exterminate; over the long run they tend to win. Second, in the U.S. at least, people tend to look at politics up close, and when you're myopic like that it appears that the government is a force onto itself. But zooming out, U.S. government actually mirrors the will of the people extremely well (with the exception of some issues on the margin). If there is overwhelming political support for redistribution it would be very difficult to resist.

The second equilibrium seems more likely-- the capitalist class grants the public a bare minimum to keep us from forcing political action. In the AI world "the minimum" is probably a much better standard of living than we have now, as the marginal cost of many products and services approaches zero. So we end up living much better material lives, but are still not free. Maybe this is stable, or maybe the ruling class loses dominance over time. At that point, who knows.

The future you're describing doesn't seem likely to me because in this event the public will force redistribution through political action.

Brass tacks, if an institution has an overwhelming political leaning toward faction X and works to undermine faction Y, is it really surprising that when Y gets into power it attempts to damage the institution? This is precisely why publicly funded institutions should maintain agnostic political posture.

I had the opposite experience. Opus 4.6 extended feels like the first genuinely intelligent model to converse with, Opus 4.7 adaptive feels like slightly smarter LinkedIn slop.

In a city not only does it do random things, when it does work it’s calibrated so poorly people behind me signal all the time because it’s too slow.

On a freeway it’s only kind of usable. It switches lanes far too aggressively and for no reason, to the point that it makes the ride uncomfortable.

What I really want is auto steer with lane switching when I signal, which for some reason I could never get working in any mode. It either doesn’t change lanes at all, or changes them arbitrarily of its own volition. And if I change lanes manually it turns off autosteer, which is too irritating to use in practice.

Tesla self driving, in any mode, is a bad product. And I say this as a Tesla fan.

A view from my small corner on the inside: taste isn't merely not incentivized, it's actively disincentivized. It's not selected for during the interview process, if you demonstrate a little of it nobody cares, if you demonstrate too much of it you clash with everyone else's priorities which quickly becomes career limiting. So people willing to fight for taste never advance.

This isn't some nefarious plot to screw over users. Taste is not prioritized because nobody has it and thus can't recognize it. Can't value something you don't even recognize. This is orthogonal to talent btw. Lots of people there who are insanely good at what they do, who produce the most hideous API specs you've ever seen, as one example.

A much more mundane (and almost certainly true) explanation is that people who put all that crap in legitimately thought it's a good idea. Taste is its own thing and it's just not in Microsoft's DNA.

Now I have no illusions about who looked stupid and who were stupid.

Could you expand on what you mean?

Incredible letters, thanks for sharing. I wish some of this correspondence was published in physical books. What a joy it would be to read.

A model that gets good at computer use can be plugged in anywhere you have a human. A model that gets good at API use cannot. From the standpoint of diffusion into the economy/labor market, computer use is much higher value.

Why do you need to keep up? Just use the latest models and don't worry about it.

I can honestly understand both positions. The U.S. military must be able to use technology as it sees fit; it cannot allow private companies to control the use of military equipment. Anthropic must prevent a future where AIs make autonomous life and death decisions without humans in the loop. Living in that future is completely untenable.

What I don’t understand is why the two parties couldn’t reach agreement. Surely autonomous murderous robots is something U.S. government has interest in preventing.

Your comment reminds me of a story. John Adams and Lafayette met in Massachusetts something like ~49 years after the revolution. (Lafayette went on a US tour to celebrate the upcoming 50 year anniversary of independence.) Supposedly after the meeting Adams said "this was not the Lafayette I knew" and Lafayette said "this was not the Adams I knew".

Yeah, it's the easiest way to get a beachhead in deep learning and then expand from there. I dislike their heavy use of currying, it's elegant in theory but bad error messages make it confusing and inconvenient in practice. But it's a small tradeoff for an otherwise excellent book.

Ah, I'm using racketmode which doesn't support live state buildup (and the builtin GUI doesn't either). What exactly is your setup? SLIME only has a Common Lisp backend, it doesn't support Racket to my knowledge.

EDIT: ok with geiser and geiser-racket incremental state buildup works really well. I rescind my objection!